A new $5T industry has emerged: AI roll-ups
Our view on AI roll-ups: skeptical of the traditional buy-then-AI model — and bullish on a net-new path that builds similar outcomes through AI-native delivery, bypassing the old factory.
Two views — both true
The eye-catching headlines put a giant number on AI roll-ups. Fair enough. The services markets in play are enormous in aggregate. The number is still not the point.
Here is our view:
We are skeptical of the traditional model — buy fragmented services businesses, then “implement AI” — for the same reasons seasoned operators in small and lower-middle-market deals keep repeating. Most of that pitch fails the real world.
We are also bullish on a net-new business model that largely bypasses that traditional path: build the same (or better) client outcome from the ground up around AI delivery, instead of buying a labor-arbitrage factory and trying to retrofit agents into it under debt and change-management pressure. The technology is not finished. It is close enough that the people who rethink the model — not just bolt AI onto the old one — will reshape industries.
Those are not contradictions. They are two different games.
Why we are skeptical of traditional buy-then-AI
Knowledge-work services businesses are largely labor-arbitrage machines. They sell human hours, packaged as expertise. Margins are thin. Scalability is constrained by headcount. Moats are often switching cost and relationship inertia, not proprietary product.
The popular AI roll-up story says: buy those businesses, compress delivery with agents, keep the clients, expand margins, roll them up. Easy to describe. Hard to execute. AI does not change that.
Seasoned operators who live in the acquisition ecosystem are right about the ground truth:
Basic tech debt comes first. A large share of SMB and lower-middle-market targets still do not run modern CRM, clean books, or documented processes. Putting AI on top of that does not create an edge. It automates chaos.
Roll-ups are people and integration problems. Key operators leave. Clients follow the human they trusted, not the holding company logo. Tribal knowledge does not live in a slide deck. Day four after close can erase the model.
Debt plus unproven AI assumptions is how you get hurt. Leverage on a thesis that agents will remove cost on a schedule you have not proven is not underwriting. It is hope with interest payments.
Tools diffuse. Once a useful AI workflow works in an industry, it tends to become reasonably widely adopted. “We use AI” is not a durable moat. Operating discipline is.
We agree with that operator view. We share it. Traditional buy-then-AI — cold-buy a services shop, invent the factory after the wire, service acquisition debt while you learn the business — is a bad sequence dressed up as a new asset class.
AI does not skip stabilize
The acquisition lifecycle is usually acquire, then stabilize, then grow, then exit. The AI margin story lives in grow. Before you get there, you have to survive stabilize.
That is the part the buy-then-AI pitch skips. You inherit employees you did not hire, customers you did not win, vendors you did not pick, processes you did not design, systems you did not choose, financials you did not prepare, and a culture you did not build. Often the institutional knowledge is still in the seller’s head — and they just got wired a life-changing check. The debt payment is still due every month.
A lot of acquirers spend years between stabilize and grow. They rebuild accounting. They fix working capital. They replace people and customers. They learn the industry. They put out fires that never showed up in the CIM. They may still create real wealth. They often never reach the clean, optimized growth phase they modeled.
AI can still improve margins. It can cut administrative labor, tighten quoting and sales process, and create operating leverage. Everyone buying or operating a business should be thinking about it. Growth still needs stability, management bandwidth, usable systems, clean data, and cash. AI is a tool. It is not a substitute for competent operations. You still have to buy the right company at the right price with the right capital structure, keep the right people, protect the customer base, manage cash, and actually operate the thing.
The hard part is getting the business into a position where you can capture the AI opportunity. AI did not make that part easier.
What the clean pitch still leaves out
Be careful with the takeaway that someone can buy a business, stand up agents after hours, and triple its profit.
You are taking on two hard jobs at once: learning to operate that business, and changing how it delivers its work. Having access to the same models as a large fund does not give you the same ability to do either.
A detailed playbook makes the process understandable. It can also make you feel prepared for situations you have never dealt with. Knowing the steps in an acquisition does not tell you whether a particular company is worth buying, why its customers stay, or how much of the business depends on the owner who is leaving.
Then you add the assumptions about AI. You assume the work is repeatable. The data is usable. Employees can explain the exceptions they have spent twenty years learning. Customers will stay through the transition. The agents will need little enough supervision to make the savings worthwhile.
Each of those needs evidence. A working demo can create confidence in the whole plan when it has only proved that one task can be completed under certain conditions.
Even when the automation works, time saved does not automatically become profit. If it creates capacity to serve more customers, you still have to win those customers. If the plan is to reduce costs, you need to establish which expenses actually disappear after implementation, review, and maintenance — not which ones disappear on the spreadsheet.
The risk that gets people hurt is borrowing money against the outcome they expect, with several untested assumptions between the purchase and that outcome. If implementation takes longer, a key employee leaves, or customers follow the former owner, you are learning to run the company while trying to fix the acquisition plan. Debt payments and payroll continue throughout.
Oversimplification is its own risk. The clean version of this story travels faster than the operating truth. It makes the opportunity look obvious and the people doing the actual work look slow. Tech founders who have never run anything that was not software are now talking about buying a boring business the way people talk about a side project. Half are “looking at a deal.” Most have never carried payroll through a bad quarter in a services firm.
Respect the operators you would buy
Surviving small and mid-market services businesses are not dumb. They have been operating through cycles, managing client relationships, and staying solvent in competitive markets for years. Walking in with an AI playbook and a thesis about how you will “fix their industry” is not a strategy — it is a posture that gets expensive fast.
The real dynamic inside most of these businesses is resource scarcity. Not enough time. Not enough capital. Not enough bandwidth to implement a CRM correctly, let alone rebuild delivery around AI agents while keeping existing clients serviced and existing debt paid.
When the owner’s kids do not want the business, that is information. They watched what it took for thirty years and said no. Treat that refusal as a diligence input, not as proof that a bargain is waiting.
Many of these firms are supply-constrained, not demand-constrained. Agents help delivery. They do not invent licenses, physical capacity, or judgment on demand. A twelve-person firm carries twelve people’s worth of context. Meeting notes are not a substitute for that.
The other path: bypass the old factory
Here is the second half of our view — the part that is easy to dismiss as crazy until it is not.
If the traditional model fails because you are buying a labor machine and trying to rewrite it under pressure, then the logical alternative is to stop buying the labor machine as the first move.
Build a net-new business whose delivery OS is designed around agents from day one: agents draft, process, categorize, flag, and prepare; humans own judgment, client trust, regulatory sign-off, and cash decisions. Sell the outcome clients already buy — reports, closes, filings, campaigns, reviews — without inheriting someone else’s tribal debt, owner dependence, and tech stack from 1998.
That is not “sprinkle AI on a roll-up.” That is a different company shape:
- Prove delivery before you buy. Run the agent-assisted model on live work until the hand-off points are real, not assumed.
- Acquire later, if at all, into a working system. Buy client books or complementary capacity into infrastructure that already functions — not invent the factory after the wire.
- Underwrite AI like an operator. Repeatable work, usable data, exception paths, supervision cost, and which expenses actually disappear must be evidence, not slides.
- Treat tools as table stakes. The durable edge is the rulebook: what agents may draft, what humans must verify, who signs off, and how cash and board truth stay honest.
Is the technology fully there for every knowledge-work vertical today? No. Parts of it are. More of it will be soon. The window that matters is not “AI works perfectly.” The window is “who redesigns the business model while everyone else argues about whether to bolt AI onto the old one.”
The people who flip the model — rethink intake, delivery, pricing, and trust for an agent-heavy stack — will reshape industries. The people who only add AI to a traditional roll-up will mostly learn expensive lessons about CRM, key-person risk, and debt.
Constraint thinking still applies
Even on the net-new path, local task automation is not system throughput. Speeding up draft production tenfold when the bottleneck is senior review just grows the queue. Time saved is not cost removed until you prove which costs leave after review and maintenance.
Net-new does not mean “no humans.” It means humans are placed on judgment, trust, and sign-off by design — not as leftovers from a business you bought and hoped would transform.
Where Vantage Rock stands
We are an AI-native financial leadership and finance-ops firm. We work with founder-led businesses, PE portfolio companies, and PE firms. We are building how finance work gets delivered in this shift — because we believe the category is moving from theory to integration fights, and because we are skeptical of theater.
We are not neutral observers. We are practitioners.
If you are a founder, the question is whether you compete on hours or whether you own a delivery system. Those are different businesses with different valuations and different risks.
If you are a PE team evaluating a portco services business, the question is whether the asset is worth rebuilding under a buy-then-AI plan — or whether you are about to overpay for a labor-arbitrage model that is structurally compressing — and whether a net-new delivery OS is the cleaner path to the outcome you actually want.
If you are an operator inside any of this, the question is what agents should draft, what processes should be systematized, and what humans must own — now, not in two years.
Our view in one line: skeptical of traditional buy-then-AI roll-ups — AI does not skip stabilize; bullish on building the next model that does not need the old factory first.
The thesis deserves weight. So do the operators who keep saying the hard part out loud. The future belongs to the people who listen to both — and still rebuild.
If you are navigating any of this, vantagerockfinancial.com.
Stavros Christias runs Vantage Rock Financial, a fractional CFO firm working with founder-led services, healthcare and multi-entity businesses. Ten-plus years across FP&A, controllership, reporting, forecasting and systems implementation, including PE-backed operators. You talk to the operator, not a sales team. LinkedIn.
15–30 minute Introduction Call.
It is a fit-check, not a sales call. We don't diagnose on the call and you don't leave with a plan. Thirty minutes tells us both whether there's work here worth doing.